Text Orientation Detection Based on Multi Neural Network
Zhiyao Zhou, Lan Lin · 2020
Optical character recognition (OCR) is an important research area in the field of pattern recognition, such as Vehicle License Plate Recognition. With it, we can extract textual information from the images to facilitate digital processing. However, most existing systems are designed to detect or recognize horizontal (or near-horizontal) texts and can't be applied to recognize texts of varying orientations. Text Orientation Detection is an important but challenging task. It can be used as a pre-processing for previous researches, and allows them to be adapted to more complex situations. Once we get angle of the text, we can use a series of transformations to get horizontal text. Although this problem is a multiclass classification, the results using common multiclass classification methods are not ideal. Our algorithm is inspired by the human behavior of recognizing texts. In this paper, we propose a new algorithm containing three neural networks to detect text orientation. The first is for capturing the abstract information of text image, and trained on multi-orientation, synthetic texts. Then the second is for evaluating the correctness of meaning of texts and trained on horizontal texts. The output of these two neural networks serves as the input to the last. In this way, the last neural network can obtain information about the image of the text as well as its meaning, and finally evaluate and output the angle of the text. According to the results, our proposed method can guarantee a high accuracy of text orientation detection.